Self-supervised knowledge distillation for few-shot single-stage 3D object detection

Conventional deep learning models for 3D object detection require a large number of annotated bounding boxes and substantial computational resources, which limit their applicability in data- and resource-constrained autonomous driving scenarios. To address this challenge, this paper proposes SSD-KD, a self-supervised single-stage 3D object detector that integrates knowledge distillation and incremental few-shot learning to improve both data efficiency and computational efficiency. SSD-KD employs a static–dynamic co-teaching strategy, where a frozen static teacher preserves previously learned knowledge while a dynamically updated teacher adapts to new data, enabling continual learning with reduced catastrophic forgetting. A channel-wise autoencoder is introduced to distill both semantic and geometric representations by compressing and reconstructing intermediate bird’s eye view features, facilitating effective feature-level knowledge transfer. Self-supervised learning enables the model to exploit a large amount of unlabeled LiDAR data, further reducing reliance on extensive manual annotations. Furthermore, incremental few-shot learning capabilities enable SSD-KD to generalize well to previously unseen data. Extensive experiments have been conducted on the KITTI and nuScenes datasets, and SSD-KD shows substantial performance gains under few-shot and cross-dataset evaluation protocols while maintaining real-time efficiency.